1 ; match686.asm -- Asm portion of the optimized longest_match for 32 bits x86
2 ; Copyright (C) 1995-1996 Jean-loup Gailly, Brian Raiter and Gilles Vollant.
3 ; File written by Gilles Vollant, by converting match686.S from Brian Raiter
4 ; for MASM. This is as assembly version of longest_match
5 ; from Jean-loup Gailly in deflate.c
8 ; http://www.winimage.com/zLibDll
9 ; http://www.muppetlabs.com/~breadbox/software/assembly.html
11 ; For Visual C++ 4.x and higher and ML 6.x and higher
12 ; ml.exe is distributed in
13 ; http://www.microsoft.com/downloads/details.aspx?FamilyID=7a1c9da0-0510-44a2-b042-7ef370530c64
15 ; this file contain two implementation of longest_match
17 ; this longest_match was written by Brian raiter (1998), optimized for Pentium Pro
18 ; (and the faster known version of match_init on modern Core 2 Duo and AMD Phenom)
20 ; for using an assembly version of longest_match, you need define ASMV in project
22 ; compile the asm file running
23 ; ml /coff /Zi /c /Flmatch686.lst match686.asm
24 ; and do not include match686.obj in your project
26 ; note: contrib of zLib 1.2.3 and earlier contained both a deprecated version for
27 ; Pentium (prior Pentium Pro) and this version for Pentium Pro and modern processor
28 ; with autoselect (with cpu detection code)
29 ; if you want support the old pentium optimization, you can still use these version
31 ; this file is not optimized for old pentium, but it compatible with all x86 32 bits
32 ; processor (starting 80386)
35 ; see below : zlib1222add must be adjuster if you use a zlib version < 1.2.2.2
37 ;uInt longest_match(s, cur_match)
39 ; IPos cur_match; /* current match */
42 cur_match equ dword ptr[esp+NbStack-0]
43 str_s equ dword ptr[esp+NbStack-4]
44 ; 5 dword on top (ret,ebp,esi,edi,ebx)
45 adrret equ dword ptr[esp+NbStack-8]
46 pushebp equ dword ptr[esp+NbStack-12]
47 pushedi equ dword ptr[esp+NbStack-16]
48 pushesi equ dword ptr[esp+NbStack-20]
49 pushebx equ dword ptr[esp+NbStack-24]
51 chain_length equ dword ptr [esp+NbStack-28]
52 limit equ dword ptr [esp+NbStack-32]
53 best_len equ dword ptr [esp+NbStack-36]
54 window equ dword ptr [esp+NbStack-40]
55 prev equ dword ptr [esp+NbStack-44]
56 scan_start equ word ptr [esp+NbStack-48]
57 wmask equ dword ptr [esp+NbStack-52]
58 match_start_ptr equ dword ptr [esp+NbStack-56]
59 nice_match equ dword ptr [esp+NbStack-60]
60 scan equ dword ptr [esp+NbStack-64]
62 windowlen equ dword ptr [esp+NbStack-68]
63 match_start equ dword ptr [esp+NbStack-72]
64 strend equ dword ptr [esp+NbStack-76]
65 NbStackAdd equ (NbStack-24)
74 ; all the +zlib1222add offsets are due to the addition of fields
75 ; in zlib in the deflate_state structure since the asm code was first written
76 ; (if you compile with zlib 1.0.4 or older, use "zlib1222add equ (-4)").
77 ; (if you compile with zlib between 1.0.5 and 1.2.2.1, use "zlib1222add equ 0").
78 ; if you compile with zlib 1.2.2.2 or later , use "zlib1222add equ 8").
82 ; Note : these value are good with a 8 bytes boundary pack structure
83 dep_chain_length equ 74h+zlib1222add
84 dep_window equ 30h+zlib1222add
85 dep_strstart equ 64h+zlib1222add
86 dep_prev_length equ 70h+zlib1222add
87 dep_nice_match equ 88h+zlib1222add
88 dep_w_size equ 24h+zlib1222add
89 dep_prev equ 38h+zlib1222add
90 dep_w_mask equ 2ch+zlib1222add
91 dep_good_match equ 84h+zlib1222add
92 dep_match_start equ 68h+zlib1222add
93 dep_lookahead equ 6ch+zlib1222add
102 public _longest_match
108 MIN_LOOKAHEAD equ (MAX_MATCH+MIN_MATCH+1)
114 MIN_LOOKAHEAD equ (MAX_MATCH + MIN_MATCH + 1)
115 MAX_MATCH_8_ equ ((MAX_MATCH + 7) AND 0FFF0h)
118 ;;; stack frame offsets
120 chainlenwmask equ esp + 0 ; high word: current chain len
122 window equ esp + 4 ; local copy of s->window
123 windowbestlen equ esp + 8 ; s->window + bestlen
124 scanstart equ esp + 16 ; first two bytes of string
125 scanend equ esp + 12 ; last two bytes of string
126 scanalign equ esp + 20 ; dword-misalignment of string
127 nicematch equ esp + 24 ; a good enough match size
128 bestlen equ esp + 28 ; size of best match so far
129 scan equ esp + 32 ; ptr to string wanting match
132 ; saved ebx byte esp + 36
133 ; saved edi byte esp + 40
134 ; saved esi byte esp + 44
135 ; saved ebp byte esp + 48
136 ; return address byte esp + 52
137 deflatestate equ esp + 56 ; the function arguments
138 curmatch equ esp + 60
140 ;;; Offsets for fields in the deflate_state structure. These numbers
141 ;;; are calculated from the definition of deflate_state, with the
142 ;;; assumption that the compiler will dword-align the fields. (Thus,
143 ;;; changing the definition of deflate_state could easily cause this
144 ;;; program to crash horribly, without so much as a warning at
145 ;;; compile time. Sigh.)
147 dsWSize equ 36+zlib1222add
148 dsWMask equ 44+zlib1222add
149 dsWindow equ 48+zlib1222add
150 dsPrev equ 56+zlib1222add
151 dsMatchLen equ 88+zlib1222add
152 dsPrevMatch equ 92+zlib1222add
153 dsStrStart equ 100+zlib1222add
154 dsMatchStart equ 104+zlib1222add
155 dsLookahead equ 108+zlib1222add
156 dsPrevLen equ 112+zlib1222add
157 dsMaxChainLen equ 116+zlib1222add
158 dsGoodMatch equ 132+zlib1222add
159 dsNiceMatch equ 136+zlib1222add
162 ;;; match686.asm -- Pentium-Pro-optimized version of longest_match()
163 ;;; Written for zlib 1.1.2
164 ;;; Copyright (C) 1998 Brian Raiter <breadbox@muppetlabs.com>
165 ;;; You can look at http://www.muppetlabs.com/~breadbox/software/assembly.html
168 ;; This software is provided 'as-is', without any express or implied
169 ;; warranty. In no event will the authors be held liable for any damages
170 ;; arising from the use of this software.
172 ;; Permission is granted to anyone to use this software for any purpose,
173 ;; including commercial applications, and to alter it and redistribute it
174 ;; freely, subject to the following restrictions:
176 ;; 1. The origin of this software must not be misrepresented; you must not
177 ;; claim that you wrote the original software. If you use this software
178 ;; in a product, an acknowledgment in the product documentation would be
179 ;; appreciated but is not required.
180 ;; 2. Altered source versions must be plainly marked as such, and must not be
181 ;; misrepresented as being the original software
182 ;; 3. This notice may not be removed or altered from any source distribution.
185 ;GLOBAL _longest_match, _match_init
190 ;;; uInt longest_match(deflate_state *deflatestate, IPos curmatch)
194 longest_match proc near
196 _longest_match proc near
198 .FPO (9, 4, 0, 0, 1, 0)
200 ;;; Save registers that the compiler may be using, and adjust esp to
201 ;;; make room for our stack frame.
207 sub esp, LocalVarsSize
209 ;;; Retrieve the function arguments. ecx will hold cur_match
210 ;;; throughout the entire function. edx will hold the pointer to the
211 ;;; deflate_state structure during the function's setup (before
212 ;;; entering the main loop.
214 mov edx, [deflatestate]
217 ;;; uInt wmask = s->w_mask;
218 ;;; unsigned chain_length = s->max_chain_length;
219 ;;; if (s->prev_length >= s->good_match) {
220 ;;; chain_length >>= 2;
223 mov eax, [edx + dsPrevLen]
224 mov ebx, [edx + dsGoodMatch]
226 mov eax, [edx + dsWMask]
227 mov ebx, [edx + dsMaxChainLen]
232 ;;; chainlen is decremented once beforehand so that the function can
233 ;;; use the sign flag instead of the zero flag for the exit test.
234 ;;; It is then shifted into the high word, to make room for the wmask
235 ;;; value, which it will always accompany.
240 mov [chainlenwmask], ebx
242 ;;; if ((uInt)nice_match > s->lookahead) nice_match = s->lookahead;
244 mov eax, [edx + dsNiceMatch]
245 mov ebx, [edx + dsLookahead]
249 LookaheadLess: mov [nicematch], ebx
251 ;;; register Bytef *scan = s->window + s->strstart;
253 mov esi, [edx + dsWindow]
255 mov ebp, [edx + dsStrStart]
259 ;;; Determine how many bytes the scan ptr is off from being
267 ;;; IPos limit = s->strstart > (IPos)MAX_DIST(s) ?
268 ;;; s->strstart - (IPos)MAX_DIST(s) : NIL;
270 mov eax, [edx + dsWSize]
271 sub eax, MIN_LOOKAHEAD
277 ;;; int best_len = s->prev_length;
279 mov eax, [edx + dsPrevLen]
282 ;;; Store the sum of s->window + best_len in esi locally, and in esi.
285 mov [windowbestlen], esi
287 ;;; register ush scan_start = *(ushf*)scan;
288 ;;; register ush scan_end = *(ushf*)(scan+best_len-1);
289 ;;; Posf *prev = s->prev;
291 movzx ebx, word ptr [edi]
293 movzx ebx, word ptr [edi + eax - 1]
295 mov edi, [edx + dsPrev]
297 ;;; Jump into the main loop.
299 mov edx, [chainlenwmask]
305 ;;; match = s->window + cur_match;
306 ;;; if (*(ushf*)(match+best_len-1) != scan_end ||
307 ;;; *(ushf*)match != scan_start) continue;
309 ;;; } while ((cur_match = prev[cur_match & wmask]) > limit
310 ;;; && --chain_length != 0);
312 ;;; Here is the inner loop of the function. The function will spend the
313 ;;; majority of its time in this loop, and majority of that time will
314 ;;; be spent in the first ten instructions.
316 ;;; Within this loop:
319 ;;; edx = chainlenwmask - i.e., ((chainlen << 16) | wmask)
320 ;;; esi = windowbestlen - i.e., (window + bestlen)
326 movzx ecx, word ptr [edi + ecx*2]
331 LoopEntry: movzx eax, word ptr [esi + ecx - 1]
335 movzx eax, word ptr [eax + ecx]
339 ;;; Store the current value of chainlen.
341 mov [chainlenwmask], edx
343 ;;; Point edi to the string under scrutiny, and esi to the string we
344 ;;; are hoping to match it up with. In actuality, esi and edi are
345 ;;; both pointed (MAX_MATCH_8 - scanalign) bytes ahead, and edx is
346 ;;; initialized to -(MAX_MATCH_8 - scanalign).
352 mov edx, 0fffffef8h; -(MAX_MATCH_8)
353 lea edi, [edi + eax + 0108h] ;MAX_MATCH_8]
354 lea esi, [esi + eax + 0108h] ;MAX_MATCH_8]
356 ;;; Test the strings for equality, 8 bytes at a time. At the end,
357 ;;; adjust edx so that it is offset to the exact byte that mismatched.
359 ;;; We already know at this point that the first three bytes of the
360 ;;; strings match each other, and they can be safely passed over before
361 ;;; starting the compare loop. So what this code does is skip over 0-3
362 ;;; bytes, as much as necessary in order to dword-align the edi
363 ;;; pointer. (esi will still be misaligned three times out of four.)
365 ;;; It should be confessed that this loop usually does not represent
366 ;;; much of the total running time. Replacing it with a more
367 ;;; straightforward "rep cmpsb" would not drastically degrade
374 mov eax, [esi + edx + 4]
375 xor eax, [edi + edx + 4]
380 LeaveLoopCmps4: add edx, 4
381 LeaveLoopCmps: test eax, 0000FFFFh
388 ;;; Calculate the length of the match. If it is longer than MAX_MATCH,
389 ;;; then automatically accept it as the best possible match and leave.
397 ;;; If the length of the match is not longer than the best match we
398 ;;; have so far, then forget it and return to the lookup loop.
400 mov edx, [deflatestate]
404 mov esi, [windowbestlen]
405 mov edi, [edx + dsPrev]
407 mov edx, [chainlenwmask]
410 ;;; s->match_start = cur_match;
412 ;;; if (len >= nice_match) break;
413 ;;; scan_end = *(ushf*)(scan+best_len-1);
415 LongerMatch: mov ebx, [nicematch]
417 mov [edx + dsMatchStart], ecx
422 mov [windowbestlen], esi
423 movzx ebx, word ptr [edi + eax - 1]
424 mov edi, [edx + dsPrev]
426 mov edx, [chainlenwmask]
429 ;;; Accept the current string, with the maximum possible length.
431 LenMaximum: mov edx, [deflatestate]
432 mov dword ptr [bestlen], MAX_MATCH
433 mov [edx + dsMatchStart], ecx
435 ;;; if ((uInt)best_len <= s->lookahead) return (uInt)best_len;
436 ;;; return s->lookahead;
439 mov edx, [deflatestate]
441 mov eax, [edx + dsLookahead]
447 ;;; Restore the stack and return from whence we came.
449 add esp, LocalVarsSize
456 ; please don't remove this string !
457 ; Your can freely use match686 in any free or commercial app if you don't remove the string in the binary!
458 db 0dh,0ah,"asm686 with masm, optimised assembly code from Brian Raiter, written 1998",0dh,0ah
472 _match_init proc near